CREDIT Research Paper

Running Title (short) 1

No. 19/11

Early childhood health during conflict: The legacy of theLord’s Resistance Army in Northern Uganda

by

Sarah Bridges and Douglas Scott

Abstract

This study finds evidence of irreversible health deficits amongst young children whowere exposed to the Lord’s Resistance Army insurgency in Northern Uganda (1987-2007). The causal effect of the conflict is found to be a 0.65 standard deviation fall inheight-for-age z-scores amongst children exposed for a period of more than six months.In contrast, the health impacts of shorter periods of exposure are found to be relativelyminimal. These findings highlight the need for a swift resolution to conflict, in par-ticular where it impacts heavily upon civilian populations, without which, the healthconsequences of protracted wars may extend far beyond the current generation.

JEL Classification: I12, J13, O12

Keywords: Conflict, Uganda, Child health

Centre for Research in Economic Development and International Trade,University of Nottingham

CREDIT Research Paper

Running Title (short) 1

No. 19/11

Early Childhood Health During Conflict: The Legacy ofthe Lord’s Resistance Army in Northern Uganda

by

Sarah Bridges and Douglas Scott

Outline

1. Introduction

2. The War in Northern Uganda

3. Data

4. Empirical Strategy

5. Results

6. Robustness

5. Heterogeneity

6. Conclusion

References

Appendices

The Authors

Corresponding Author: douglas.scott1@nottingham.ac.uk

Sarah Bridges (sarah.bridges@nottingham.ac.uk) and Douglas Scott are respectivelyAssociate Professor and Teaching Fellow at the University of Nottingham, School ofEconomics.

Acknowledgements

Thanks go to Oliver Morrissey for advice throughout this research

Research Papers at www.nottingham.ac.uk/economics/credit/

1 Introduction

Few conflicts on the continent of Africa have gained as much notoriety as the twenty-

year war fought between the government forces of Uganda and the Lord’s Resistance

Army (LRA). Under the leadership of the self-proclaimed spiritual medium Joseph

Kony, the shockingly brutal tactics employed by the LRA against the civilian popu-

lation drew international condemnation. Although attempts have been made to analyse

the post-war recovery of those caught up in the conflict (Blattman, 2009; Blattman and

Annan, 2010; Bozzoli and Bruck, 2010; Adelman and Peterman, 2014; Fiala, 2015),

there exists no direct study of the impact of the war on the physical health of the civil-

ian population. This paper exploits spatial and temporal variation in the spread of the

conflict to measure the causal effect of the war on the health outcomes of children liv-

ing in the affected districts. Evidence is found of irreversible health deficits for children

exposed to the conflict for a period of more than 6 months, with children in this group

experiencing an average shortfall in height-for-age z-scores of 0.65 standard deviations.

In contrast, however, there is no evidence of significant height deficits amongst those

exposed to the fighting for shorter periods of time. These results are found to be robust

to alternative samples, aimed at addressing potential sources of bias, and alternative

definitions of conflict exposure, with a similar pattern of results also observed in other

anthropometric measures of health status. Given potential links between early child-

hood health and outcomes in later life (Strauss and Thomas, 2007), the deficits expe-

rienced by these children are likely to impact the economic prospects of war-affected

regions for many years to come.1

The remainder of this section presents a review of the relevant literature to which

this study contributes. Section 2 provides a brief history of the LRA conflict which took

1See Alderman et al. (2006) on malnutrition in Zimbabwe and completed grades of education, Mac-cini and Yang (2009) on the impact of rainfall shocks on health, education and asset wealth in Indonesia,and Dercon and Porter (2014) on the long-run health and income effects experienced by survivors of the1984 Ethiopian famine.

1

place in northern Uganda, while a description of the data used, and the characteristics of

the sample, can be found in Section 3. The empirical strategy employed to estimate the

causal effect of the conflict on health outcomes is outlined in Section 4, with the main

results of this analysis reported in Section 5. A comprehensive study of robustness and

potential sources of heterogeneity can be found in Section 6 and 7, before a summary

of the key findings and concluding remarks are presented in Section 8.

1.1 Previous Studies

This paper makes a contribution to three specific strands of the literature. Firstly, the

analysis adds to the growing number of studies which endeavour to quantify the impact

of conflict on early childhood health. For example, Bundervoet et al. (2009) provide evi-

dence of a cumulative effect of exposure to the 1993-2005 Burundian civil war on a sam-

ple of children aged between 6 months and 5 years. Each additional month of exposure

to the war is found to decreases children’s height-for-age z-scores by 0.047 standard

deviations, with an average differential between exposed and non-exposed children of

-0.348 to -0.525. Akresh et al. (2012b) use a similar methodology to estimate the effect

of the 1998 Eritrean-Ethiopian war on children’s health, yielding comparable results.

Their findings indicate that children alive during the war, and living in a war-affected

region, would be between 0.447 and 0.454 standard deviations shorter than those who

were not exposed to the conflict. More recently, Minoiu and Shemyakina (2014) show

similar results for children exposed to the 2002-2007 civil war in the Cote d’Ivoire.2

Secondly, these results contribute to the literature looking specifically at the ef-

fects of the LRA insurgency in northern Uganda. Amongst these studies, Blattman and

Annan (2010) analyse the labour market outcomes of children who were previously

2Other notable contributions on the impact of conflict on children’s health can be found in Akreshet al. (2011), which covers the civil war preceding the 1994 Rwandan genocide, Shemyakina (2017), inrelation to politically motivated violence in Zimbabwe, and Akresh et al. (2012a) who focus on the heightdeficits of adults born during the Nigerian civil war of 1967-1970.

2

abducted by the rebels to serve as child-soldiers.3 The authors estimate the loss of hu-

man capital from time spent away from education and employment, coupled with the

psychological distress of increased exposure to violence, led to 33% lower wages for

former abductees. Much attention has also been given to the economic consequences of

the widespread displacement caused by the conflict. For example, Fiala (2015) finds ev-

idence of a negative impact on both consumption and asset wealth amongst previously

displaced households, with only wealthier households showing later signs of recovery.

In another analysis of conflict-driven displacement, Adelman and Peterman (2014) find

that resettled households experienced significant losses in access to agricultural land

upon return to their former locations. Evidence also exists of less tangible impacts of

the LRA conflict. For example, Rohner et al. (2013) find an increase in ethnic identity

following the fighting, hampering economic recovery in more fractionalised communi-

ties, while Bozzoli et al. (2011) show that exposure to the conflict reduces individual’s

expectations of their future life situation and economic prosperity. Links have also been

found between exposure to violence and increased political participation amongst ex-

combatants (Blattman, 2009), although this may only take place locally, as a response

to the immediate needs placed upon communities, rather than as a result of an increased

concern over politics at the national level (De Luca and Verpoorten, 2015).

Finally, this study contributes to the literature on gender bias in childhood health

outcomes and the potential for heterogeneity in the impact of negative shocks. For

example, Baird et al. (2011) analyse a sample of 59 low-income countries and find

relatively higher infant mortality amongst female children in response to fluctuations

in per-capita GDP, while Rose (1999) finds that positive rainfall shocks increase the

probability of survival for girls in rural India (relative to male children). There is,

however, little evidence of a gender bias in the small number of papers that specifically

3It is estimated that between 60,000 and 80,000 children and young adults were abducted during theconflict, mostly from the northern, Acholi districts, which were formerly Kitgum and Gulu (Blattmanand Annan, 2010).

3

address childhood health outcomes in response to conflict shocks. For example, Minoiu

and Shemyakina (2014) uncover no evidence of significant heterogeneity in height-for-

age z-scores of male and female children who were exposed to conflict in Cote d’Ivoire.

Similarly, Akresh et al. (2011) do not find evidence of a gender bias as a result of the

Rwandan civil war, in spite of a clear bias towards male children in response to crop

failures in other parts of the country. A more recent study by Dagnelie et al. (2018)

does find lower survival rates among female children during the 1997-2003 civil war in

the Democratic Republic of Congo. However, the authors attribute this to an adverse

selection effect, driven by the lower probability of survival for male children in utero,

rather than a gender bias influencing their subsequent survival prospects.

2 The War in Northern Uganda

As with many other violent disputes in East Africa, the origins of Uganda’s LRA con-

flict can be traced to the long-standing political and ethnic divisions within the coun-

try. Historically, Uganda was divided between the predominantly Bantu South, and the

Nilotic (Nilo-Hamitic) and Central Sudanic people of the North and Northwest (Rohner

et al., 2013). Following independence from British rule in 1962, the North provided the

majority of Uganda’s military power, paving the way for a series of brutal Northern dic-

tatorships, which would serve to concentrate political power within the hands of those

loyal to the current head of state. In 1986 this dominance ended with the defeat of the

de facto government forces (largely comprised of the Northern Acholi and Langi ethnic

groups) by the National Resistance Army (NRA) led by Yoweri Museveni, a South-

erner and veteran of the campaign which ousted the notorious dictator Idi Amin (Allen,

2013). Following the NRA victory, defeated Northern soldiers retreated back to their

4

homelands, with many crossing the border into southern Sudan. From the remnants of

these forces a number of armed resistance groups initially formed in opposition to the

new government. By 1988 most of these groups had either signed peace agreements

with Museveni’s government or been defeated by forces loyal to the new regime. The

decision to stop fighting was not unanimous, however, and a small number of soldiers,

who were unwilling to accept the outcome of peace negotiations, gathered under the

leadership of Joseph Kony, in what would become the Lord’s Resistance Army (Allen,

2013).

Kony belonged to the Acholi ethnic group and it was the districts of Gulu and Kit-

gum (known as Acholi-land) which would initially become the centre of the LRA in-

surgency.4 There existed little local support for the poorly financed rebel group prior

to 1995, making them almost entirely reliant on the abduction and forced conscription

of civilians (often children) to maintain their numbers (Blattman, 2009). Although the

LRA’s stated aim was to overthrow the Ugandan government, increasing Kony began to

target civilian populations, proclaiming that Acholi society must be ‘purified’ to over-

come its oppressors (Doom and Vlassenroot, 1999; Allen, 2013). Following an initially

slow campaign, the number of violent attacks started to escalate in 1995, when the LRA

received support from the Sudanese government, in the form of arms, provisions and

land to establish bases in southern Sudan (Dolan, 2009). Attacks on civilians increased

notably during this period, occurring seemingly at random, and with overwhelming

brutality (Doom and Vlassenroot, 1999; Blattman and Annan, 2010). Many of the rural

inhabitants of Acholi-land and neighbouring districts sought protection closer to local

towns or military outposts and from 1996 the government began forcibly relocating the

4These two districts were later subdivided into seven districts (Amuru, Nwoya, Gulu, Agago,Lamwo, Pader and Kitgum). However, for clarity and consistency with the empirical analysis whichfollows, district names are recorded as they stood in 1991, which corresponds to the earliest birth yearrecorded for children in the studied sample (see Section 3.1).

5

Acholi population to internally displaced people (IDP) camps situated in these loca-

tions. The camps were intended to shield civilians from the LRA attacks, but in reality,

they offered little protection. Instead, they were characterised by overcrowding, poor

sanitation and an abundance of disease (Bozzoli and Bruck, 2010).

During the late 1990s, the plight of northern Uganda was gaining international at-

tention, placing pressure on the Sudanese government to cease support for the LRA.

In 2002 the Ugandan military (with support from the US) obtained permission from

Khartoum to launch operation ‘Iron Fist’, a military incursion against the LRA bases in

southern Sudan. However, Kony and almost all senior LRA members survived the raid

and the rebels outflanked the government forces, attacking new territories in Lira, Apac

and Soroti (Allen, 2013). The failure of this operation marked a rapid increase in the

number of attacks and fatalities attributed to the LRA, with the period between 2002 and

2005 constituting the height of the violence experienced during the campaign (Rohner

et al., 2013; De Luca and Verpoorten, 2015). By 2004 approximately 1.5 million people

were estimated to be internally displaced as a result of the fighting (International Crisis

Group, 2004) and increased global awareness of this humanitarian crisis, along with

the numerous atrocities committed by the LRA, led the International Criminal Court to

issue an arrest warrant against Kony and four of his senior commanders (Dolan, 2009).

With global attention fixed on the conflict, pressure was placed on both sides to reach

a peaceful solution. In mid-2006 the government of Uganda agreed to engage in peace

talks with the rebels, resulting in the signing of a cessation of hostilities agreement in

Juba (Sudan) on 26th August 2006. Although Kony never signed the final peace agree-

ment, these talks began the effective end of the LRA war in northern Uganda (Dolan,

2009).

6

3 Data

3.1 Health and Demographic Data

The individual-level health data used in this analysis comes from three waves of the

Uganda Demographic and Health Survey (DHS), collected in 1995, 2000 and 2006.5

Data is provided on a number of key childhood health indicators, including height-for-

age, weight-for-age and weight-for-height. The DHS surveys also contain demographic,

health and education measures relating to the child’s mother and the characteristics of

the household. In order to accurately link this data to specific locations and events, the

sample considered only includes observations for children whose mothers were present

in the same DHS sample cluster since the child’s birth.6 This initial sample consists

of a cross-section of 9496 children, born between 1991 and 2006, and aged less than

5 years old at the time of the survey.7 The geographical coverage of the three surveys

varies to some degree, as does the ages of the children selected for height and weight

measurements (children were only measured before 48 months in 1995 survey). How-

ever, the identification strategy employed, coupled with evidence from estimations on

alternative samples (Section 6.2), suggests that the main findings of this analysis are not

unduly influenced by this.

The primary outcome of interest is the height-for-age z-score of children within the

sample, which measures the number of standard deviations from the (age and gender-

specific) median height of a child in a healthy reference population.8 This is a long-run

5At the time of writing, all Uganda DHS data used in this study is available on request from dhspro-gram.com/data.

6The likelihood of bias in the estimated impact of conflict on health, due to a correlation betweenhousehold relocation and health status, is discussed in Section 6.1.

7However, the main analysis is conducted using only a sub-sample of 9202 children who were borneither before or during the fighting (see Section 3.3).

8The impact of conflict exposure on other anthropometric measures is considered in Section 6.6.

7

measure of exposure to poor health and nutrition, implying that an accumulation of past

health deficits should still be visible in the data, even amongst older children.9

Figure 1: Height-for-Age z-scores by Month of Birth

-3-2

-10

1H

eigh

t-for

-age

z-s

core

DHS 1995 DHS 2000 DHS 2006

Month of Birth and DHS Survey Dates

The curves in Figure 1 illustrate the relationship between height-for-age z-scores

and a child’s month of birth, generated using locally weighted, scatter-plot smoothing

(LOWESS). Each separate curve represents data points from a different DHS survey,

with the start and finish dates of the three data collections shown by the vertical lines

in Figure 1. In each case, the far left of the curve represents the oldest children at the

time of survey, while the z-scores of younger children will be represented by the area

where the curve is closest to (or within) the time period where the survey took place.

Figure 1 clearly shows that each curve slopes upward towards the dates of the respective

survey, indicating that as children in the sample grow older their height lags increasingly

9This measure is also indicative of an individual’s future health and economic status. An extensiveliterature has found strong links between early childhood height and physical and cognitive development,morbidity, mortality, schooling and economic productivity in later life. See Strauss and Thomas (2007)for a review.

8

behind that of the reference population.10 As only around 1 in 10 children in the sample

were directly exposed to the fighting, this suggests that, even in the absence of conflict,

children born in Uganda during this period are likely to have faced substantial health

challenges.

3.2 Conflict Data and Localities

Using the sub-county locations of the DHS clusters and the children’s dates of birth,

health outcomes are linked to information on LRA conflict events obtained from the

Armed Conflict Location Events Database (ACLED).11 For the purposes of this analy-

sis, conflict events are defined as any recorded battle or act of one-sided violence, where

the LRA is listed as one of the actors. The ACLED dataset records 1947 such events,

occurring between January 1987 and March 2007.

The twenty-year duration of the LRA insurgency presents clear challenges for iden-

tification. For example, no children in the DHS data were measured prior to the first

event taking place.12 A second concern relates to the geographical proximity of the

children’s households to the recorded locations of where events actually took place.13

Therefore, to mitigate these concerns, while fully exploiting the spatial and temporal

variation of the fighting, Uganda is sub-divided into 309 localities, along county and

subcounty administrative boundaries. These localities are constructed by sub-dividing

any county, where the distance between two border locations exceeds 50kms, along

10A cumulative deficit in height-for-age, especially before 3 years of age (see Fig 1), conforms with along-recognised pattern in low-income countries (Martorell and Habicht, 1986).

11Events occurring from the 1st of January 1997 onwards are taken from the current version ofACLED, accessed from https://www.acleddata.com/data/ on 17th June 2019 (Raleigh et al., 2010).Events taking place before this date are obtained from an earlier version of the same data, compiledby the Peace Research Institute Oslo (Raleigh and Hegre, 2005).

12One earlier DHS survey was conducted in Uganda in 1988/89. However, other than the West Nileregion (Arua, Moyo and Nebbi districts in Fig 2), the survey was only conducted in the south and south-west of the county.

13Studies using a similar methodology have classified children living more than 100kms from anyconflict event as exposed to the fighting. For example, see Akresh et al. (2012b) and Minoiu and She-myakina (2014).

9

subcounty administrative boundaries (if possible). For counties/municipalities where

this distance never exceeds 20kms, the area is combined with an adjoining county.14 A

description of the areas covered by the localities can be found in Appendix 1.

Figure 2: Number of Recorded LRA Conflict Events, by Locality

GULUKITGUM

MUKONO

KOTIDO

MOROTOLIRA

IGANGA

ARUA

SOROTI

APAC

MBARARA

LUWERO

MASINDI

MPIGI

MASAKA

MOYO

HOIMA

KABAROLE

RAKAI

MUBENDE

KAMULI

KUMI

KIBAALE

NEBBI

KASESE

KALANGALA

KIBOGA

BUSHENYI

MBALE

TORORO

RUKUNGIRI

PALLISA

KABALE

BUNDIBUGYO

KAPCHORWA

JINJA

KISORO

KAMPALA

0 100 20050 Kilometers

Events1 - 56 - 2021 - 5051 - 100101 - 200201 - 288

Source: Based on ACLED dataset (Raleigh and Hegre, 2005; Raleigh et al., 2010)

14One exception to this approach occurs in the case of the geographically small area of Gulu munici-pality, which occupies a location directly between Aswa and Omoro counties. In this instance, the threecounties/municipalities are combined, and then subdivided along subcounty lines thereafter.

10

Figure 2 shows the extent of the LRA conflict, which took place between 1987 and

2007, along with the sub-division of Uganda’s 38 districts (as of 1991) into the 309

localities. Darker shaded areas represent localities that experienced a greater number of

events, based on the ACLED dataset (Raleigh and Hegre, 2005; Raleigh et al., 2010).

The highest intensity of fighting can clearly be observed in the Northern Acholi and

Langi sub-regions (Kitgum, Gulu, Apac and Lira), as well as the district of Soroti,

further to the southeast.

3.3 Exposure to Conflict

The initial approach to assigning whether or not a child is considered as being exposed

to the fighting is based on the methodology used by Bundervoet et al. (2009) and re-

quires first defining a conflict window for each locality where at least one event took

place. This conflict window represents the time period between (and including) the

calendar month where the first and last recorded events occurred in the locality. The

simplest definition of exposure assigns any child who was alive during the conflict win-

dow as exposed to the conflict, while omitting from the sample all children who were

born after the conflict window. Initial estimations will, therefore, only utilise observa-

tions from children who were either exposed to the conflict or measured prior to any

fighting taking place. In the analysis which follows, a number of alternative defini-

tions of conflict exposure are considered, both in the initial estimations and by way

of robustness checks in Section 6. The table provided in Appendix 2 provides a more

detailed description of the 49 localities in the sample which experienced at least one

event, henceforth, referred to as conflict localities.

Columns 1 to 3, in Table 1, compare the characteristics of children in conflict and

non-conflict localities, whereas columns 4 to 6 report the same variables for exposed

and non-exposed children, within conflict localities. One of the clearest disparities lies

11

in the difference between the average heights of children and their mothers, across the

two sets of results. Surprisingly, Table 1 suggests that, on average, mothers and chil-

dren are actually taller in the areas where the fighting took place (relative to areas which

were unaffected). Furthermore, children who were exposed to the fighting, within these

areas, appear significantly taller than those who were not exposed. This seemingly

counterintuitive finding can be understood by recognising that the majority of the fight-

ing took place in the districts of Gulu, Kitgum, Lira, Apac and Soroti (see Household

Characteristics in Table 1), where the Nilotic ethnic groups, which dominate these ar-

eas, have far closer historical links to the famously tall Dinka people of southern Sudan

or the Maasai of Kenya and northern Tanzania than those in the south and southwest of

Uganda (Shoup, 2011). On closer examination of Table 1, it is also clear that, within

the observations from conflict localities, children exposed to the fighting have a higher

probability of coming from one of these five districts (see columns 4 to 6). Therefore,

the over-representation of Nilotic children in the exposed group (within the conflict ar-

eas) would also potentially explain the significant height (and possibly weight) disparity

in column 6.15

To formally test the likelihood of underlying anthropometric differences between

ethnic groups, Appendix 3 reports an analysis of the relationship between ethnicity and

height, conducted on a sample of women who would have achieved adult stature before

the LRA insurgency began in 1987. This simple empirical analysis confirms that adult

women from the groups which dominate the five most affected districts have signifi-

cant height advantages over those in other conflict localities and, indeed, elsewhere in

Uganda. A naıve analysis which does not recognise this would (at best) understate the

health impacts of the LRA insurgency on those who were most affected by the conflict.

15For example, repeating the tests reported in Table 1, column 6, with the addition of district fixedeffects, fully negates any significant difference in height or weight, between exposed and non-exposedchildren (height-for-age p-value = 0.730, weight-for-age p-value = 0.606).

12

Table 1: Descriptive Statistics

(1) (2) (3) (4) (5) (6)

Locality Conflict Exposure

Non- Diff. Not Diff.Variable Conflict Conflict (1)-(2) Exposed Exposed (4)-(5)

Mean Mean p-value Mean Mean p-value

Child’s CharacteristicsHeight-for-age z-score −1.34 −1.55 0.065* −1.19 −1.57 0.009***Weight-for-age z-score −0.99 −1.11 0.134 −0.87 −1.18 0.010**Height-for-weight z-score −0.19 −0.18 0.818 −0.16 −0.24 0.228Child is female 0.52 0.51 0.347 0.51 0.54 0.202Child age (years) 2.14 2.04 0.030** 2.24 1.98 0.012**Birth order 1-3 0.46 0.46 0.853 0.44 0.49 0.155Birth order 4-6 0.32 0.33 0.608 0.35 0.29 0.042**Birth order 7+ 0.22 0.20 0.506 0.21 0.22 0.761

Mother’s CharacteristicsMother’s age (years) 28.54 28.17 0.190 28.99 27.85 0.036**Mother’s height (cms) 160.85 157.87 0.000*** 161.18 160.36 0.379Mother is married/cohabiting 0.91 0.91 0.909 0.91 0.91 0.902Mother has no education 0.31 0.26 0.207 0.26 0.39 0.036**Mother has primary education 0.58 0.61 0.235 0.61 0.54 0.150Mother has secondary education 0.11 0.13 0.596 0.13 0.08 0.232Mother is head 0.90 0.92 0.118 0.92 0.87 0.135Mother is daughter of head 0.07 0.05 0.046** 0.06 0.08 0.329Mother not head/daughter 0.03 0.03 0.639 0.02 0.05 0.139Mother’s time in location 13.04 11.79 0.155 11.96 14.66 0.053*

Household CharacteristicsUrban area 0.20 0.18 0.844 0.30 0.04 0.011**Gulu, Kitgum, Lira, Apac, Soroti 0.61 0.02 0.000*** 0.66 0.53 0.487Household size (time of survey) 6.78 6.54 0.196 6.62 7.03 0.337Household head is male 0.82 0.83 0.485 0.80 0.85 0.015**Household head’s age (years) 36.20 35.64 0.194 35.96 36.55 0.485

Observations 1447 7755 869 578

Test standard errors clustered at the locality * p<0.1, ** p<0.05, *** p<0.01

The measures in Table 1 also suggest significant variation in the mean ages of chil-

dren between groups, with evidence of relatively older children present in conflict local-

ities and also in the group who were exposed to the fighting. In the latter case, a higher

13

probability of being exposed to the war, amongst relatively older children, would almost

certainly be a consequence of having a longer period of time in which an event can oc-

cur, whereas the fact that this group are older, on average, would account for a higher

birth order and also mothers who are relatively older when the survey is conducted. The

reason behind the difference in the mean ages of children in conflict and non-conflict

localities is not initially so clear, however. One explanation is that this may be related

to the differences in the sampling methodology in the three DHS survey waves. In the

1995 wave, only children who were younger than 4 provided height and weight mea-

surements (as opposed to children younger than 5 in the 2000 and 2006 surveys). The

1995 survey also did not cover the district of Kitgum, where all but one of the 17 lo-

calities were affected by the insurgency. Therefore, in 1995, both older children and

conflict localities are relatively under-sampled.

A significant difference in the mean ages of children between the groups in Table 1

clearly presents a possible threat to identification.16 However, it is important to recog-

nise that the identification strategy outlined in the following section addresses this by

only estimating the effects of the LRA conflict using variation within birth cohorts.

Similarly, all variables which show significant heterogeneity between the groups in Ta-

ble 1 (such as the gender of the household head or the education level of the mother)

are added as additional controls in a separate set of estimations, to assess the effect of

their inclusion on the main coefficients of interest.17

16Assuming relatively older children will have accumulated larger height deficits (see Figure 1), themeasured effect of conflict exposure would be downward biased if this disparity in ages was not acknowl-edged in the identification strategy outlined in Section 4.

17Differences in the educational attainment of the children’s mothers may reflect a relatively highernumber of children in urban areas who were exposed to the fighting (see 1). Repeating the tests reportedin Table 1, controlling for rural/urban status, produces p-values in excess of 0.25 for all education groups.

14

4 Empirical Strategy

To attempt to measure the causal effect of the war on the height-for-age z-scores of

affected children, three alternative definitions of conflict exposure are considered. The

baseline model (1) follows Bundervoet et al. (2009) in measuring the impact of the

LRA insurgency on children who were both living in affected localities and exposed

to the conflict. Given a higher relative mean age of exposed children, coupled with

possible differences in the underlying anthropometrics of those in conflict and non-

conflict localities (see Appendix 3), the following baseline specification identifies the

effect of exposure to the war, using only variation within localities and within birth

cohorts.

hazijt = αj + δt + β1(Conf Localityj ∗ Exposedi) + φ1femalei + εijt (1)

In model (1), hazijt represents the height-for-age z-score of child i, living in locality

j, who was born in year t, while the terms αj and δt represent locality and birth cohort

(year of birth) fixed effects. Conf Locality takes value 1 if the child comes from an

area which experienced at least one conflict event (0 otherwise) and the Exposed term

in this baseline model is simply a binary indicator for whether or not a child was alive

during the conflict window. In the first set of estimations, only a control for the child’s

gender is included. The final term ε in model (1) denotes a random, idiosyncratic error.

One potential drawback of model (1) is that only a relatively small number of con-

flict localities contain both exposed and non-exposed children (see Appendix 2). Given

the inclusion of the locality fixed effect αj , the coefficient of interest β1 will only be

estimated using variation within this small subset of localities. Furthermore, due to the

longer duration of the fighting in the worst affected districts (resulting in fewer pre-

war observation) the areas where variation in the binary measure of exposure exist will

15

commonly be found on the peripheries of the most intense fighting.18 In light of this,

model (2) replaces Exposed with the variable Ex Duration, measuring the number of

months a child was alive during the conflict window. In (2) the coefficient β2 measures

the effect of an additional month of exposure, under the implicit assumption of a linear

relationship between months of exposure and childhood health. All other terms in (2)

have the same interpretation as before.

hazijt = αj + δt + β2(Conf Localityj ∗ Ex Durationi) + φ1femalei + εijt (2)

A third definition of exposure is shown in model (3), which relaxes the assumption

that the marginal health impact of an additional month in the conflict window is uniform

across all possible exposure durations. In (3), children in affected localities who were

exposed to the violence are assigned to one of two groups, dependant on the time they

were alive during the conflict window. The coefficients β3 and β4 measure the effect

of exposure to the fighting in a conflict locality (for the given durations), relative to

those children who were not directly exposed to the war. The initial estimation of (3)

separates the exposed sample at a conflict duration of 6 months, while in the results in

Section 5 also defined the two groups around a split point of 12 months.

hazijt = αj + δt + β3(Conf Localityj ∗ Ex 1− 6monthsi)

+β4(Conf Localityj ∗ Ex > 6monthsi) + φ1femalei + εijt

(3)

18The mean duration of the conflict windows in localities from the worst affected districts (Gulu, Kit-gum, Lira, Apac and Soroti) is approximately 110 months, compared to only around 35 months outside ofthese districts. Furthermore, a relatively lower percentage of the sampled children in these districts werenot exposed to the fighting (only 16.4%, as opposed to 39.7% elsewhere, based on the table in Appendix2).

16

In each of the previous models, identification relies on the assumption that the un-

derlying time trend in children’s heights would be the same in both conflict and non-

conflict localities, had the insurgency not taken place. This may be an overly strong

requirement, given the focus of the fighting in the Northern districts and the large eco-

nomic, cultural and ethnic divide between the North and South of the country. Fol-

lowing the literature, therefore, this condition can be modified by the assumption that

children in conflict and non-conflict localities, within districts, would display similar

growth pattern in the absence of exposure. This requires the inclusion of a district-

specific time trend in models (1) to (3), which is indicated in the case of the baseline

model below by the term γjt.

hazijt = αj + δt + γjt + β1(Conf Localityj ∗ Exposedi) + φ1femalei + εijt (4)

The main results of this analysis also report estimations for each of the three defini-

tions of exposure including a number of additional controls, relating to the child, their

mother and the household in which they live. Following a discussion of the robustness

of findings from the previous specifications, Section 7 also extends the main results to

assess heterogeneity in the effect of conflict between different sub-sample groups. This

section focuses on three potential sources of heterogeneity, the intensity of the fighting

experienced by the child and the gender of either the head of the household or the child

themselves.

17

5 Results

Table 2 reports results from estimations based on the models described in the previ-

ous section. The first three columns show measures of the β1 interaction coefficient

from the baseline model (1) (all terms in (1) are included in the estimations, but not

reported in Table 2). Under the assumption of a common time trend in height-for-age

z-scores, column 1 reports little impact of the war on children residing in conflict lo-

calities and exposed to the fighting (relative to those not directly affected). With the

inclusion of the district-specific time trends in column 2, however, there is evidence of

a 0.37 standard deviation height deficit amongst those who were exposed. Table 1 sug-

gests several underlying characteristics may differ systematically between children in

conflict and non-conflict localities (as well as between exposed and non-exposed chil-

dren). After the inclusion of variables capturing the characteristics of the child, their

mother, and the household,19 column 3 indicates the estimate of β1 is only minimally

affected, suggesting that the causal effect of exposure to the war may be closer to -0.39.

The magnitude of the coefficients in columns 2 to 3 could be considered somewhat

small, relative to comparable studies.20 As previously noted, however, the localities

which contain both exposed and non-exposed children tend to be found on the edges of

the worst fighting.21 As an alternative to the simple (yes/no) exposure variable, columns

4 to 6 report the effect of an additional month of exposure to the conflict, measured by

β2 in model (2). Again, significant coefficients are found in the estimations includ-

ing district-specific time trends, with both column 5 and 6 reporting a 0.011 standard

deviation deficit in height for each additional month spent within the conflict window.

19An alternative version of Table 2, including the estimated coefficients on all additional controls, isprovided in Appendix 4.

20Bundervoet et al. (2009) found a -0.348 to -0.525 standard deviation height deficit in their study ofthe 1993-2005 civil war in Burundi, Akresh et al. (2012b) reported coefficients of -0.447 to -0.454 forthe border war between Eritrea and Ethiopia, while Minoiu and Shemyakina (2014) found an effect ofexposure to conflict between -0.250 and -0.414 (for the Cote d’Ivoire).

21For example, only a single locality in the Acholi-land districts of Gulu and Kitgum provides varia-tion in the initial definition of exposure, while these two districts account for 1384 of the recorded events(71.1% of the total 1947 events).

18

Table 2: The Effects of Conflict Exposure on Height-for-Age

Dep. Variable:height-for-age z-score (1) (2) (3) (4) (5) (6)

Conf Locality * Exposed -0.239 -0.369** -0.388**(0.187) (0.186) (0.195)

Conf Locality * Ex Duration 0.001 -0.011** -0.011**(0.003) (0.005) (0.005)

Observations 9202 9202 9202 9202 9202 9202R2 0.079 0.147 0.205 0.079 0.147 0.205Locality Fixed Effect Yes Yes Yes Yes Yes YesBirth Cohort Fixed Effect Yes Yes Yes Yes Yes YesDistrict-Specific Trend No Yes Yes No Yes YesControls† No No Yes No No Yes

(7) (8) (9) (10) (11) (12)

Conf Locality * Ex 1-6m -0.191 -0.200 -0.234(0.213) (0.215) (0.218)

Conf Locality * Ex > 6m -0.315* -0.659*** -0.651***(0.165) (0.194) (0.222)

Conf Locality * Ex 1-12m -0.231 -0.298 -0.317(0.197) (0.201) (0.204)

Conf Locality * Ex > 12m -0.258 -0.566*** -0.585**(0.171) (0.200) (0.227)

Observations 9202 9202 9202 9202 9202 9202R2 0.079 0.148 0.206 0.079 0.147 0.205Locality Fixed Effect Yes Yes Yes Yes Yes YesBirth Cohort Fixed Effect Yes Yes Yes Yes Yes YesDistrict-Specific Trend No Yes Yes No Yes YesControls† No No Yes No No Yes

H0: Ex 1-6(12)m = Ex > 6(12)mp-value 0.293 0.005*** 0.020** 0.710 0.083* 0.096*

Locality clustered standard errors in parentheses * p<0.1, ** p<0.05, *** p<0.01

In (7) to (9): Not exposed (n = 578), 1-6m exposure (n = 218), exposure > 6m (n = 651)In (10) to (12): Not exposed (n = 578), 1-12m exposure (n = 344), exposure > 12m (n=525)†Child controls include: female child and birth order (grouped)Mother’s controls include: height, age, marital and education status, and relationship to head.Household controls include: urban DHS survey cluster, age and gender of the headThe full table of results is shown in Appendix 4

19

Columns 7 to 9 report results from estimations based on model (3). This approach splits

the sample of children into groups, based on whether a child’s duration of exposure is

less than (equal to) or more than 6 months. The results in columns 8 and 9 show clear

evidence of a greater deficit for those children who were exposed to the fighting for

more than 6 months, with the results suggesting this may be in excess of -0.65 standard

deviations (relative to those not affected). In contrast, it is not possible to reject the

hypothesis of no impact for those exposed over shorter periods of time (at any of the

levels reported). In columns 8 and 9, a Wald test of the equality of the impact of different

exposure durations also confirms heterogeneity in the negative effects experienced by

the two groups (p-values reported in Table 2).

The final set of results in columns 10 to 12 splits the sample of exposed children

at 12 months, as opposed to 6. The results appear broadly similar to those in columns

7 to 9, although the estimated effect of inclusion in the longer exposure group appears

less negative. There also exists weaker evidence of a significant difference between

the shorter and longer periods of exposure when separating observations at 12 months

duration (see p-values in Table 2). This pattern of results suggests an acceleration in the

negative health effects of exposure to the LRA conflict between 6 and 12 months, and

minimal evidence of further detrimental effects thereafter.

With no variance in the binary exposure measure in the majority of localities from

the worst affected districts, the models which exploit variation in the length of exposure

are preferable. Having split the sample of affected children at 6 months exposure, Table

2 also suggests little additional insight can be gained from also splitting the sample at 12

months. Therefore, the remainder of the analysis will focus attention on the estimations

used to generate the results in column 6 and column 9 of Table 2 (with both additional

controls and district-specific time trends).

20

6 Robustness

This section considers the robustness of the results reported in Table 2 (columns 6 and 9)

to the use of samples intended to address potential sources of bias, alternative definitions

of conflict localities (and the conflict window), and the use of alternative anthropometric

health measures.

6.1 Endogenous Selection out of Conflict Exposure

Arguably, the clearest source of potential bias comes from an endogenous selection

process governing which children are exposed to conflict. For example, if relatively

wealthier households were more easily able to relocate to safer areas, and in the likeli-

hood that there exists a positive correlation between the health status of children and the

economic status of households, the results in Table 2 will be biased (downwards). Un-

fortunately, there exists no information on household migration in the Ugandan DHS

surveys. However, this information is recorded in the Ugandan National Household

Survey 2005/06 (UNHS), which was collected just prior to the 2006 DHS wave. The

UNHS records whether household members moved to their current location since 2001

and provides information on the district of origin and their main reason for relocation.

Of the 2749 households in the UNHS, who either resided in one of the 16 conflict-

affected district or moved from one of these areas, 283 had relocated with the stated

reason of avoiding insecurity (see Appendix 5). The vast majority of these households

(89.4%) had settled within the same district, however, implying that the bulk of conflict-

driven relocation would have occurred at a relatively local level.22 If households with

22A summary of the migration pattern of households from the affected districts, taken from the UgandaNational Household Survey 2005/06, is provided in Appendix 5, where the migration history of thehousehold is assumed to follow the relocation history of the household head.

21

healthier children were able to leave conflict areas, either before the fighting started

or at a relatively earlier stage, there should exist a negative correlation between time

in the current location and the child’s height, within localities yet to be affected by

the fighting. Testing for this correlation indicates that the relationship between time in

location and height-for-age z-scores is found to be significantly negative (p< 0.05) only

in three districts, Gulu, Hoima and Kotido.23 To establish the extent of any possible bias

in the results reported in Table 2 (columns 6 and 9), these estimations are repeated on a

sample which omits observations from the three districts listed above. The results can

be found in columns 1 and 2 of Table 3.

Comparing the estimated effects in Table 3 to the initial results suggest that the

linear exposure duration effect (in column 1) may not be wholly robust to the omission

of the three districts above, although the magnitude of the effect is only reduced slightly.

However, in the case of the grouped results in column 2, any downward bias in the

measured effects appears to be relatively small and certainly does not warrant any re-

interpretation of the findings.24

6.2 Variation in Survey Coverage between Waves

As the population of interest is Ugandan children aged less than 60 months, two further

potential sources of bias can be found in the districts selected for the DHS surveys and

the ages of children measured in these surveys. The 1995 wave only measured children

who were aged less than 4 years old and did not survey the Northern district of Kitgum.

23A description of the estimation used to test for a negative correlation between time in location andheight-for-age z-scores can be found in Appendix 6.

24The reason that selection effects appear negligible may be linked to the seemingly random nature ofthe LRA attacks (Blattman and Annan, 2010; Adelman and Peterman, 2014). Furthermore, in the case ofrelocation to IDP camps, this was often involuntary, with little or no notice, making endogenous selectionout of conflict exposure even less likely (Fiala, 2015).

22

In contrast, the survey conducted in 2000 omitted both Kitgum and Gulu, along with

the districts of Kasese and Bundibugyo, while increasing the age range to 5 years of

age.25 Similarly, the final 2006 survey measured children up to 5 years old, but in this

final wave, all districts of Uganda were surveyed.

The districts which were not surveyed in either 1995 or 2000 were omitted due to

fears of insecurity, implying an obvious correlation between the probability of inclusion

in the original DHS sample and conflict exposure. This suggests the estimated impact of

the LRA conflict in Table 2 may contain a positive bias (Children in the worst affected

areas were not surveyed). The absence of children aged 4 or over in the 1995 sample

implies a further potential selection effect, although in this case, the sign of the bias is

less clear.26 To address these concerns, the key models are re-estimated on a sample

of children which is consistent (both in age and geographical coverage) across all three

DHS survey waves.

The results in columns (3) and (4) of Table 3 show the impact of the war on chil-

dren aged less than 4 years, and not living in the districts of Gulu, Kitgum, Kasese

and Bundibugyo. This alternative sample should be fully representative of a popula-

tion with these characteristics. Unsurprisingly, given that this sub-sample omits over

1200 observations and considers children over a different age range, the magnitude of

the coefficients varies to some degree (relative to those in column 6 and 9 of Table 2).

However, the fact that the broad pattern of results remains unchanged is further reassur-

ance that any sample selection bias is not sufficient to warrant a new interpretation of

the previous results.

25The latter two Western districts were experiencing insecurity during this period, due to another rebelgroup ADF-NALU operating along the border of Uganda and the Democratic Republic of Congo.

26For example, the relative under-sampling of older children in 1995, and the likelihood of cumulativehealth deficits from longer periods of exposure, would be expected to generate a positive bias. However,if the effects of a given length of exposure are more detrimental to children at a younger age, the biaswould be in the opposite direction.

23

Table 3: Alternative Samples and Assumptions on Exposure

(1) (2) (3) (4) (5) (6)

Dep. Variable: Gulu, Hoima and Districts and Ages Re-classify Conflictheight-for-age z-score Kotido Omitted Consistent all DHS Localities in South

Conf Locality * Ex Duration -0.009 -0.015** -0.011**(0.005) (0.007) (0.005)

Conf Locality * Ex 1-6m -0.220 -0.135 -0.369(0.224) (0.246) (0.303)

Conf Locality * Ex > 6m -0.614*** -0.598*** -0.742***(0.234) (0.223) (0.269)

Observations 8889 8889 7973 7973 9202 9202R2 0.200 0.201 0.211 0.212 0.205 0.206Locality Fixed Effect Yes Yes Yes Yes Yes YesBirth Cohort Fixed Effect Yes Yes Yes Yes Yes YesDistrict-Specific Trend Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes Yes Yes

H0: Ex 1-6m = Ex > 6mp-value 0.040** 0.011** 0.055*

(7) (8) (9) (10) (11) (12)

Children in IDP Extending Conflict Extending ConflictCamps Omitted Window + 6m Window + 12m

Conf Locality * Ex Duration -0.009* -0.012** -0.014***(0.005) (0.005) (0.005)

Conf Locality * Ex 1-6m -0.213 -0.079 0.165(0.228) (0.183) (0.163)

Conf Locality * Ex > 6m -0.638*** -0.617*** -0.611***(0.232) (0.186) (0.171)

Observations 9091 9091 9244 9244 9291 9291R2 0.205 0.206 0.204 0.205 0.204 0.205Locality Fixed Effect Yes Yes Yes Yes Yes YesBirth Cohort Fixed Effect Yes Yes Yes Yes Yes YesDistrict-Specific Trend Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes Yes Yes

H0: Ex 1-6m = Ex > 6mp-value 0.036** 0.002*** 0.000***

Locality clustered standard errors in parentheses * p<0.1, ** p<0.05, *** p<0.01

24

6.3 Isolated Conflict Localities

Three conflict localities in Fig 2 stand out as being much farther South than the districts

where the vast majority of the fighting took place. The localities in Bushenji, Mubende

and Jinja, experienced only five events in total during the entirety of the LRA war.

Arguably, it may not be appropriate to treat children exposed to these more isolated

pockets of violence as experiencing a continuation of the main conflict in the North.27

Therefore, as an alternative definition of exposure to conflict, the results reported in

columns 5 and 6 of Table 3 are derived from re-classifying all children in these areas as

not exposed to the conflict.

Table 3 indicates that the effect of the continuous measure of exposure (column 5)

remains unchanged. However, the more negative coefficients in column 6 suggest that

the inclusion of the children in these Southern localities (in the original estimations)

may have served to moderate the estimated impact of the war. With the limited number

of events in these areas, exposed and unexposed children (the base category in column

6) are likely to be more similar, relative to those in areas more central of the conflict,

decreasing the measured effect. Although this clearly leads to a change in the magnitude

of the coefficients, again, the overall interpretation of the results remains the same.

6.4 The Health Impacts of IDP Camps

Given the dire conditions experienced by those in the government-sanctioned IDP camps

(WHO, 2005; Bozzoli and Bruck, 2010), it initially seems plausible that the negative

health impacts attributed to the war are, in reality, capturing the consequences of chil-

dren being subjected to these conditions during displacement. Were this the case, the

27In the case of Jinja, in particular, where the two events which took place occurred more than 15years apart, many of those children categorised as exposed to the conflict will have had relatively littledirect experience of the fighting.

25

same detrimental health effects could potentially be observed anywhere where poor

sanitation and disease were prevalent (even in the absence of armed conflict).

Within the final sample, only 111 children were measured while residing in an IDP

camp, making it unlikely that these observations are responsible for determining the

results in Section 5. However, this small number of observations does highlight that

a relatively larger number of IDP children were not included in the sample due to the

mother arriving in the DHS cluster location after the child was born.28 If the impact of

conflict exposure would have been worse for these children, the results in Table 2 will

underestimate the true effect of the war. Therefore, to establish the magnitude of any

bias, while ruling out the possibility that the remaining IDP observations exert undue

influence on the results, columns 7 and 8 of Table 3 report estimates obtained from a

sample which omits IDP children.

When considering the grouped lengths of exposure in column 8 of Table 3, the es-

timated health effects of the conflict, on children who had never lived in IDP camps,

are similar to those obtained from the full sample. Where the impact of exposure dura-

tion on children’s heights is assumed to be linear, column 7 indicates a slightly smaller

effect, relative to Table 2. Overall, the results suggest that any positive bias, due to a

higher probability of IDP children being dropped from the sample, is minimal. Fur-

thermore, it is clear that the remaining observations from the IDP camps are not solely

responsible for driving the results in Section 5.

6.5 Extending the Conflict Window

In the empirical strategy defined in Section 4, the negative effects of being exposed to

the war are assumed to be limited to the period between the calendar months in which

the first and last events take place. Although this simplifying assumption avoids the

need to place an arbitrary limit on how long after the conflict window children’s health

28From the original 205 observations of IDP children, 63 were omitted due to the mother relocatingsince the child’s birth (30.7%), while this reason accounted for only around 19% of non-IDP childrendropped from the original DHS data.

26

may still be affected, it risks understating the full impact of the fighting. To determine

the sensitivity of results to alternative assumptions regarding the duration of exposure,

the conflict window is extended by 6 months after the last event in columns 9 and 10 of

Table 3, and by 12 months in columns 11 and 12.29

The linear effects reported in columns 9 and 11 appear to show an increasingly

negative impact of an additional month of fighting, when extending the limit for expo-

sure beyond the original conflict window. These results would suggest that the initial

measure of 0.011 (column 6 in Table 2) may represent a conservative estimate of the

cumulative effect of the conflict. The results in columns 10 and 12, instead, show evi-

dence of a reduction in the estimated impacts for both grouped durations of exposure.

This will come as a result of shifting the distribution of exposure duration upwards,

such that some children born after the conflict (and dropped in the initial estimation)

will now fall within the 1 to 6 month category, while some of those in this group will

now be considered as exposed for a longer duration of time. Although the estimated

effects of more than 6 months exposure are smaller in both columns 9 and 11, the re-

sults still imply a height deficit in excess of 0.6 standard deviations, with no significant

impact for shorter periods of exposure.

6.6 Alternative Anthropometric Health Measures

Height-for-age z-scores are intended to detect evidence of long-term poor nutrition and

exposure to disease. However, they represent only one potential anthropometric mea-

sure of health status available in the DHS data. If the height deficits uncovered in

Section 5 are, indeed, measuring the causal effect of conflict exposure on childhood

health, evidence should exist in more short-term measures of health status, in particular

29As children who were born after the last event were dropped from the original sample, extendingthe conflict window also increases the sample size used in these estimations.

27

amongst children who were surveyed while the fighting was still taking place. Table 4

presents an alternative set of estimations of the impact of the war on children’s health,

where the results shown in columns 3 to 6 replace the dependent variable with measures

of short-run health status, based on children’s weight (as opposed to their height). As

these measures are intended to capture short term health deficits, observations coming

from children who were measured after the conflict window are dropped from the sam-

ple for these estimations. This should serve to limit the possibility of any unrelated

health shocks impacting a child’s weight during the post-conflict period.

Table 4: Alternative Measures of Health Status

(1) (2) (3) (4) (5) (6)

Dep. Variable: Height-for-Age Weight-for-Age Weight-for-Heightz-score z-score z-score

Conf Locality * Ex Duration -0.013** -0.016*** -0.007**(0.006) (0.004) (0.003)

Conf Locality * Ex 1-6m 0.219 0.065 -0.251(0.236) (0.249) (0.211)

Conf Locality * Ex > 6m -0.580*** -0.947*** -0.672***(0.215) (0.193) (0.180)

Observations 8978 8978 8978 8978 8978 8978R2 0.206 0.207 0.192 0.195 0.125 0.126Locality Fixed Effect Yes Yes Yes Yes Yes YesBirth Cohort Fixed Effect Yes Yes Yes Yes Yes YesDistrict-Specific Trend Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes Yes Yes

H0: Ex 1-6m = Ex > 6mp-value 0.000*** 0.000*** 0.022***

Locality clustered standard errors in parentheses * p<0.1, ** p<0.05, *** p<0.01

As the results in Table 4 are estimated using a subset of the original sample, the first

two columns represent identical estimations to columns 6 and 9 of Table 2 (including the

original height-for-age measure of health status as the dependent variable). The pattern

of results is broadly similar to that in Table 2, although the estimated coefficients on the

28

grouped exposure terms are less negative (even positive in the shorter exposure group).

This is likely due to a shift in the age distribution resulting from dropping those children

measured after the conflict window. This approach will favour the inclusion of younger

children in the two exposure groups, who should have experienced shorter durations of

conflict, on average.30

The coefficients in columns 3 and 4 report estimates of the effect of conflict exposure

on weight-for-age z-scores, which measure the deviation of a child’s weight from that

of the median child (of the same age and gender) in the reference population. Under the

linear assumption in column 3, the results indicate that each additional month of expo-

sure reduces a child’s weight by 0.016 standard deviations, while the grouped duration

coefficients in column 4 still show little evidence of health impacts for an exposure du-

ration of 6 months or less. For children exposed to the conflict for more than 6 months,

however, the estimated weight deficit is remarkably large at 0.947 standard deviations,

indicating even greater heterogeneity in health impacts between the two groups, when

using weight-for-age as a measure of health status.

The final set of estimations employ weight-for-height as the dependent variable,

which records the deviation of a child’s weight from the median child, of the same

height, in the reference population. The results in columns 5 and 6 follow the same pat-

tern as those in the previous estimations. Again, there is evidence of a significant deficit

in weight for each additional month of exposure in column 5, while a strong negative

impact is found in column 6, but only for those exposed for more than 6 months. In

light of these results and those in columns 3 and 4, the similarity in the interpretation

of effects obtained using weight, as opposed to height, as an indication of health status,

provides further evidence in favour of the findings in Section 5.

30The mean age of children exposed to the war in the original sample was 26.9 months, yet this fallsto 23.7 months with the omission of those measured after the conflict window.

29

7 Heterogeneity

Having established to what extent the main results in columns 6 and 9 of Table 2 are

robust to alternative assumptions, this section considers heterogeneity in the estimated

impacts of the war on childhood health status. Three potential sources of variation in the

effect of the conflict are considered, heterogeneity between male and female children,

heterogeneity between children belonging to male and female-headed households and

heterogeneity over the intensity of the fighting experienced.

7.1 Heterogeneity by Child’s Gender

To test for evidence of a difference in the effect of the conflict between male and female

children, this study follows Minoiu and Shemyakina (2014) by interacting the conflict

exposure terms in models (2) and (3) with a female indicator variable. As it is most

likely that any gender bias would favour male children (Baird et al., 2011; Dagnelie

et al., 2018), a negative coefficient on this interaction term, when included in model (2),

would be evidence of an additional month of exposure to the war exerting a relatively

higher toll on the health of female children, than males. Similarly, a negative coefficient

on the interaction of the female indicator variable, with either of the grouped exposure

terms in model (3), would signal inclusion in these groups implied relatively worse

health outcomes for female children.

Table 5 reports results for the two sets of estimations, based on the models used

to generate the results in columns 6 and 9 of Table 2. These results are shown in the

first two columns of the table. Although the interaction terms are negatively signed in

column 1 and the longer duration group in column 2, the results show no evidence of

any additional effects of exposure to conflict amongst female children.31

31The absence of any evidence of a gender bias in conflict-related health outcomes mirrors the findingsof Akresh et al. (2011) and Minoiu and Shemyakina (2014).

30

Table 5: Heterogeneous Effects of Conflict Exposure

Dep. Variable:height-for-age z-score (1) (2) (3) (4) (5) (6)

Conf Loc * Ex Duration -0.010* -0.011** -0.011*(0.006) (0.005) (0.005)

Conf Loc * Ex 1-6m -0.238 -0.177 -0.204(0.278) (0.210) (0.298)

Conf Loc * Ex > 6m -0.630*** -0.668*** -0.612**(0.236) (0.220) (0.244)

Conf Loc * Ex Duration * Female Child -0.002(0.003)

Conf Loc * Ex 1-6m * Female Child 0.010(0.222)

Conf Loc * Ex > 6m * Female Child -0.036(0.121)

Conf Loc * Ex Duration * Female Head 0.001(0.004)

Conf Loc * Ex 1-6m * Female Head -0.328(0.259)

Conf Loc * Ex > 6m * Female Head 0.111(0.121)

Conf Loc * Ex Duration * Intensity -0.003(0.006)

Conf Loc * Ex 1-6m * Intensity -0.054(0.239)

Conf Loc * Ex > 6 months * Intensity -0.193(0.217)

Observations 9202 9202 9202 9202 9202 9202R2 0.205 0.206 0.202 0.203 0.202 0.203Locality Fixed Effect Yes Yes Yes Yes Yes YesBirth Cohort Fixed Effect Yes Yes Yes Yes Yes YesDistrict-Specific Trend Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes Yes Yes

Locality clustered standard errors in parentheses * p<0.1, ** p<0.05, *** p<0.01

7.2 Heterogeneity by the Gender of the Household Head

It is possible that gender bias in the impact of the war may operate, not through the

gender of the child, but through the gender of the head of the household. For example,

if limited resources available during conflict were disproportionately under the control

31

of males, or women experience increased vulnerability to crime as a result of the break-

down of law and order (Byrne, 1996), children within female-headed households may

suffer additional health consequences not shared by those in households where the head

was male.

Columns 3 and 4 of Table 5 report the interaction of measures of conflict exposure

with the gender of the household head. In both sets of estimations, it is not possible to

discern any significant health disadvantages amongst children in female-headed house-

holds. Taken in conjunction with the results in columns 1 and 2, these findings suggest

that the magnitude of effects were not defined by the gender of those affected.

7.3 Heterogeneity by the Intensity of Fighting

The final results in Table 5 report estimates intended to determine whether the negative

health impacts of the war are weighted towards children in areas where the intensity

of the conflict was greatest. The measure used to represent the intensity of the fight-

ing experienced is the average number of days where a conflict event took place (event

days), across all months where the child was alive during the conflict window.32 Inter-

acting this intensity measure with the conflict exposure term in column 5 generates a

coefficient which measures how the health impact of an additional month of exposure

changes, as a result of a child experiencing one more event day per month exposed. In

the grouped duration model (column 6), the coefficients on the interaction terms mea-

sure how an additional event day per month modifies the effects of inclusion in either

of the two groups.

32Based on this approach, the mean value of conflict intensity is 0.3 event days per month exposed,although the median is only 0.03 event days. Skewing of the distribution towards lower conflict intensityoccurs due to 47% of those children born within the conflict window experienced a conflict intensity of 0,implying that all events occurred outside of the period between the month they were born and the monththey were surveyed.

32

In a similar manner to the gender interactions in the previous sections, there is no

evidence of significant heterogeneity, associated with the intensity of the fighting expe-

rienced (at the levels shown in Table 5). Instead, these results suggest negative health

effects are predominantly determined by length of exposure, as opposed to the concen-

tration of events within this time period.

8 Conclusion

This study exploits spatial and temporal variation in the spread of the LRA insurgency

in northern Uganda to estimate the causal effect of exposure to conflict on the health

status of children. Linking data on 9202 children, aged between 0 to 5 years old, to the

locations of 1497 LRA conflict events, evidence is found of irreversible health deficits

amongst those exposed to the fighting for more than 6 months. The main results of

the analysis imply that, on average, children within this group display height-for-age

z-scores which are 0.65 standard deviations lower than those who were not directly

impacted by the war. Furthermore, amongst children in this group (aged 0 to 5) a deficit

in excess of 0.6 standard deviations is found using a number of alternative samples and

definitions of exposure.33 When considering the height deficits amongst the sample of

children impacted by the war for shorter periods of time, the negative effects are found

not to be significantly worse than those experienced by children who were unaffected

by the fighting.34

33Interestingly, one such alternative sample found that results were robust to the omission of obser-vation from IDP camps, suggesting camp conditions did not generate additional long-run health deficits,over and above those observed for children who were exposed to the fighting elsewhere.

34The main results also suggest a cumulative deficit in height, of 0.11 standard deviations, for eachadditional month of exposure, although the linear effect is found not to be robust to the omission ofdistricts where endogenous selection out of conflict exposure is more likely.

33

The heterogeneity in health impacts between affected children in the two duration

of exposure groups can also be observed in more short-run anthropometric measures of

health status. Estimates employing both weight-for-age and weight-for-height measures

follow an identical pattern amongst children measured in areas still exposed to the fight-

ing. The negative effects of more than 6 months exposure on weight-for-age z-scores

are particularly alarming, suggesting an average deficit approaching 1 standard devi-

ation, relative to those who were not directly exposed to the conflict. Heterogeneous

effects within the different groups of exposure duration appear largely absent, however,

with no evidence in the sample of any gender bias in the negative health impacts of the

war, either between male and female children or those living in male of female-headed

households, and no significant heterogeneity observed amongst children who experi-

enced different intensities of fighting. While the absence of such heterogeneity implies

the negative health effects of the conflict were not disproportionately weighted towards

one specific group, it also suggests that no group was capable of protecting themselves

from these effects better than any other.

The extensive literature on the long-run negative impacts of early childhood health

shocks suggests the indirect effect of the LRA insurgency may be felt in the loss of

health, education and economic wellbeing of those affected by the conflict at an early

age. Although ex-post interventions aimed at rebuilding the war-affected communities

show promising results (Blattman et al., 2013, 2016), the findings of this study, instead,

draw attention to the need for policies aimed at ensuring a swifter resolution to such

long-running conflicts. Indeed, it appears likely that the irreversible health deficits of

such events can be largely avoided if civilian populations can be returned to a state of

stability within a relatively short period of time. In the absence of a timely resolution,

however, communities affected by armed violence may be forced to carry the burden of

war, long after the last shots have been fired.

34

References

Adelman, S. and Peterman, A. (2014). Resettlement and gender dimensions of land

rights in post-conflict northern Uganda. World Development, 64:583–596.

Akresh, R., Bhalotra, S., Leone, M., and Osili, U. O. (2012a). War and stature: growing

up during the Nigerian Civil War. American Economic Review, 102(3):273–77.

Akresh, R., Lucchetti, L., and Thirumurthy, H. (2012b). Wars and child health: Ev-

idence from the Eritrean–Ethiopian conflict. Journal of development economics,

99(2):330–340.

Akresh, R., Verwimp, P., and Bundervoet, T. (2011). Civil war, crop failure, and child

stunting in Rwanda. Economic Development and Cultural Change, 59(4):777–810.

Alderman, H., Hoddinott, J., and Kinsey, B. (2006). Long term consequences of early

childhood malnutrition. Oxford economic papers, 58(3):450–474.

Allen, T. (2013). Trial justice: The international criminal court and the Lord’s Resis-

tance Army. Zed Books Ltd.

Baird, S., Friedman, J., and Schady, N. (2011). Aggregate income shocks and infant

mortality in the developing world. Review of Economics and statistics, 93(3):847–

856.

Blattman, C. (2009). From violence to voting: War and political participation in

Uganda. American political Science review, 103(2):231–247.

Blattman, C. and Annan, J. (2010). The consequences of child soldiering. The review

of economics and statistics, 92(4):882–898.

Blattman, C., Fiala, N., and Martinez, S. (2013). Generating skilled self-employment in

developing countries: Experimental evidence from Uganda. The Quarterly Journal

of Economics, 129(2):697–752.

Blattman, C., Green, E. P., Jamison, J., Lehmann, M. C., and Annan, J. (2016). Re-

turns to microenterprise support among the ultrapoor: A field experiment in postwar

Uganda. American Economic Journal: Applied Economics, 8(2):35–64.

35

Bozzoli, C. and Bruck, T. (2010). Child morbidity and camp decongestion in post-war

Uganda.

Bozzoli, C., Brueck, T., and Muhumuza, T. (2011). Does war influence individual

expectations? Economics letters, 113(3):288–291.

Bundervoet, T., Verwimp, P., and Akresh, R. (2009). Health and civil war in rural

Burundi. Journal of human Resources, 44(2):536–563.

Byrne, B. (1996). Towards a gendered understanding of conflict. IDS bulletin,

27(3):31–40.

Dagnelie, O., De Luca, G. D., and Maystadt, J.-F. (2018). Violence, selection and infant

mortality in Congo. Journal of health economics, 59:153–177.

De Luca, G. and Verpoorten, M. (2015). Civil war and political participation: Evi-

dence from Uganda. Economic Development and Cultural Change, 64(1):113–141.

Dercon, S. and Porter, C. (2014). Live aid revisited: long-term impacts of the 1984

Ethiopian famine on children. Journal of the European Economic Association,

12(4):927–948.

Dolan, C. (2009). Social torture: the case of northern Uganda, 1986-2006, volume 4.

Berghahn Books.

Doom, R. and Vlassenroot, K. (1999). Kony’s message: a new koine? the Lord’s

Resistance Army in northern Uganda. African affairs, 98(390):5–36.

Fiala, N. (2015). Economic consequences of forced displacement. The Journal of

Development Studies, 51(10):1275–1293.

International Crisis Group (2004). Northern Uganda: Understanding and solving the

conflict.

Maccini, S. and Yang, D. (2009). Under the weather: Health, schooling, and economic

consequences of early-life rainfall. American Economic Review, 99(3):1006–26.

36

Martorell, R. and Habicht, J.-P. (1986). Growth in early childhood in developing coun-

tries.

Minoiu, C. and Shemyakina, O. N. (2014). Armed conflict, household victimization,

and child health in Cote d’Ivoire. Journal of Development Economics, 108:237–255.

Raleigh, C. and Hegre, H. (2005). Introducing ACLED: An armed conflict location

and event dataset (selection of ACLED maps), paper presented at the conference

on disaggregating the study of civil war and transnational violence, University of

California Institute of Global Conflict and Cooperation, San Diego, CA.

Raleigh, C., Linke, A., Hegre, H., and Karlsen, J. (2010). Introducing ACLED: an

armed conflict location and event dataset: special data feature. Journal of peace

research, 47(5):651–660.

Rohner, D., Thoenig, M., and Zilibotti, F. (2013). Seeds of distrust: Conflict in Uganda.

Journal of Economic Growth, 18(3):217–252.

Rose, E. (1999). Consumption smoothing and excess female mortality in rural India.

Review of Economics and statistics, 81(1):41–49.

Shemyakina, O. N. (2017). Political violence and child health: Results from Zimbabwe.

World Development.

Shoup, J. A. (2011). Ethnic groups of Africa and the Middle East: An encyclopedia.

ABC-CLIO.

Strauss, J. and Thomas, D. (2007). Health over the life course. Handbook of develop-

ment economics, 4:3375–3474.

WHO (2005). Health and mortality survey among internally displaced persons in Gulu,

Kitgum, and Pader districts, northern Uganda.

37

Appendices

Appendix 1

Figure A1: Overview of Localities

67

46

7

84

68

6

139

93

103

130

9

109

167

194

30

5

81

91

23

7088

4

3

98

39

12

173

2 164

62

4086

102

1

106

24

136

166

14

64

74

32

87

187

47

22158

82

33

159

66

65

13

220

89

37

117

71

52

90

190

63

95

29

211

119

41

28

96

127

99

135

34

49

45

92

118

212

116

61

156

43

80

48

131

201

181

186

218

154

53

193

38

738572

8

155

59

210

132

157

196

216

222

179

27

121

133

77

20

202

94

227

149

228

160

57

224115

19

219 225

123

55

11

111

105

126

60

161

83

182

100

15

44172

22612858

197

75184

214

10

141

188

129

185

97

170

138223

124

35

192

5142

16

140

36

137

110

165

134

198

122

78

144

101

207

17

215

191

120

178

217

205

21

148

25

142

151

150

213

69

108

171

195

152

54

31

146

113

169107

189

56

180

20676

18

177

125

163

209

162

221

229

112

176

203

114

50

79

200

208153

204

147

168

26

145

183174

104

143

199

175

GULU

KITGUM

MUKONO

KOTIDO

MOROTOLIRA

ARUA

SOROTIMASINDI

MOYO

HOIMA

RAKAI

KUMI

IGANGA

APAC

MBARARA

LUWERO

MPIGI

MASAKA

KABAROLE

KALANGALA

MUBENDE

KAMULIKIBOGA

BUSHENYI

KIBAALE

NEBBI

KASESE

MBALE

TORORO

RUKUNGIRI

KABALE

BUNDIBUGYO

KAPCHORWA

KISORO0 100 20050 Kilometers

38

Table A1: Overview of Localities

District Locality District Locality District Locality

1 Apac Chawente 78 Tingey County 155 Rugaaga2 Kole North 79 Kasese Bukonjo North 156 Ruhaama County3 Kole South 80 Bukonjo South 157 Rwampara County4 Kwania North 81 Busongora North 158 Sanga5 Maruzi North 82 Lake Katwi 159 Moroto Checkwii East6 Oyam North 83 Muhokya 160 Iriri7 Oyam South 84 Kibaale Bugangaizi West 161 Karita8 Arua Ajia 85 Buyaga East 162 Lorengedwat9 Aringa South 86 Buyaga West 163 Lotome10 Ayivu County and Arua Mn 87 Buyanga West 164 Matheniko South/Moroto Mn11 Beleafe 88 Kiboga Kiboga Central 165 Nabilatuk12 Kei 89 Kiboga East 166 Namalu13 Koboko County 90 Kisoro Kisoro District 167 Rupa14 Madi-Okollo South 91 Kitgum Aruu Central 168 Moyo Adropi15 Maracha County 92 Awere 169 Obongi South16 Odupi 93 Chua Central 170 West Moyo County East17 Rigbo 94 Chua West 171 West Moyo County West18 Terego West 95 Lira - Palwo 172 Mpigi Busiro North19 Vurra South 96 Lokung 173 Busiro South and Entebbe Mn20 Bundibugyo Bwamba North 97 Padibe 174 Butambala East21 Harugali 98 Paimol 175 Butambala West22 Bushenyi Buhweju County 99 Parabongo 176 Gomba East23 Bunyaruguru Central 100 Patongo 177 Kabulasoke24 Igara County 101 Kotido Abim 178 Kyadondo North25 Kajara County 102 Alerek 179 Maddu26 Katerera 103 Jie North 180 Mawokota North27 Ruhinda County 104 Kaabong 181 Mawokota South28 Rushenyi County 105 Kapedo 182 Mubende Bukuya29 Sheema County 106 Kathile 183 Busujju East30 Gulu Amuru 107 Lolelia 184 Buwekula Central31 Anaka (Payira) 108 Nakaperimoru 185 Buwekula North32 Aswa/Omoro East 109 Napore (Karenga) 186 Kasambya33 Aswa/Omoro South 110 Sidok (Kopoth) 187 Kassanda South34 Aswa/Omoro West/Gulu Mn111 Kumi Bukedea North 188 Kitenga35 Bobi 112 Bukedea South 189 Maanyi36 Lamogi 113 Ngora North 190 Mityana County37 Pabbo 114 Ngora South 191 Mukono Bbaale South38 Hoima Bugahya East 115 Lira Dokolo North 192 Buikwe North39 Bugahya North 116 Erute South and Lira Mn 193 Buikwe South40 Buhanguzi East 117 Kioga East 194 Kome Islands41 Buhanguzi West 118 Moroto County East 195 Mukono North42 Iganga Bugweri County 119 Otuke East 196 Mukono South43 Bukooli North 120 Luwero Bamunanika South 197 Nakifuma Central44 Bukooli West 121 Baruli Central 198 Ntenjeru County45 Bunya North 122 Kakooge 199 Seeta46 Bunya South 123 Kalungi 200 Nebbi Jonam South47 Busiki County 124 Katikamu North 201 Okoro County48 Buyinja 125 Katikamu South 202 Padyere Central49 Kigulu County 126 Kikyusa(Kamira) 203 Pallisa Budaka County50 Luuka North 127 Nakaseke Central 204 Butebo County51 Luuka South 128 Nakaseke South 205 Kibuku County52 Jinja Jinja District 129 Masaka Bukomansimbi County 206 Pallisa County Central53 Kabale Ndorwa County/Kabale Mn 130 Bukoto East and Masaka Mn 207 Rakai Kabula North54 Rubanda East 131 Bukoto West 208 Kabula South55 Rubanda West 132 Kalungu County 209 Kooki North56 Rukiga County 133 Lwemiyaga County 210 Kooki South57 Kabarole Bunyangabu County 134 Mawogola South 211 Kyebe58 Burahya Central/Fort Portal 135 Mijwala 212 Kyotera County59 Kahunge 136 Masindi Budongo 213 Rukungiri Bujumbura South60 Kamwenge 137 Buruuli Central 214 Bwambara61 Kitagwenda County 138 Kibanda South 215 Kinkiizi North62 Kyaka North 139 Kiryandongo 216 Kinkiizi South63 Mwenge North 140 Mutunda 217 Rubabo County64 Mwenge South 141 Mbale Bubulo County 218 Soroti Kaberamaido County65 Mwenge West 142 Budadiri County 219 Kalaki County66 Nkoma 143 Bulambuli South 220 Kasilo County67 Kalangala Kyamuswa East 144 Bungokho County/Mbale Mn 221 Orungo68 Mazinga 145 Manjiya County 222 Soroti County North69 Kampala Kampala District/Makindye 146 Mbarara Buremba 223 Soroti County South/Soroti Mn70 Kamuli Budiope East 147 Ibanda Central 224 Usuk71 Budiope West 148 Ibanda North 225 Usuk County West72 Bugabula East 149 Isingiro North 226 Tororo Bunyole County73 Bugabula West 150 Isingiro South 227 Kisolo (West Budama) County74 Bulamogi County 151 Kashari East/Mbarara Mn 228 Samya - Bugwe County75 Buzaaya County 152 Kashari West 229 Tororo County and Tororo Mn76 Kapchorwa Binyini 153 Kashongi77 Kongasis County 154 Kazo Central

Localities in Figure A1 but not listed above, were not surveyed in the DHS waves used in this study

39

Appendix 2

Table A2: A Description of Conflict Localities

Conflict Window Exposed Months ExposedTotal

District Locality Name Obs. Start End Months Yes No Mean s.d.

Apac Kole North 48 2002m7 2006m6 48 5 43 2.54 9.31Apac Kwania North 44 2005m1 2005m1 1 6 38 0.14 0.35Apac Maruzi North 33 1994m4 2006m4 145 33 0 24.00 16.08Apac Oyam North 28 1987m7 2005m3 213 28 0 25.75 17.83Apac Oyam South 104 2002m3 2005m3 37 10 94 2.09 6.99Gulu Amuru 1 2005m2 2006m8 19 1 0 17.00 .Gulu Anaka (Payira) 9 1998m11 2006m8 94 9 0 31.11 18.33Gulu Aswa/Omoro East 9 1998m6 2005m7 86 5 4 11.78 12.72Gulu Aswa/Omoro South 7 1998m11 2006m2 88 7 0 20.29 20.96Gulu Aswa/Omoro West/Gulu Mn 52 1988m1 2006m8 224 52 0 20.12 12.29Gulu Bobi 3 2000m5 2005m10 66 3 0 12.67 3.51Gulu Lamogi 5 1998m9 2006m7 95 5 0 29.00 16.00Gulu Pabbo 16 1987m4 2005m2 215 16 0 18.13 14.14Kitgum Aruu Central 11 2001m4 2006m7 64 11 0 24.91 13.67Kitgum Awere 5 1998m4 2004m12 81 5 0 15.00 15.28Kitgum Chua Central 9 1996m4 2006m8 125 9 0 22.00 11.94Kitgum Chua West 7 1990m5 2007m3 203 7 0 15.14 9.96Kitgum Lira - Palwo 5 1998m4 2005m12 93 5 0 24.40 16.53Kitgum Lokung 7 2004m4 2005m7 16 7 0 12.43 3.46Kitgum Padibe 3 1991m5 2005m5 169 3 0 25.00 11.53Kitgum Parabongo 7 1997m2 2006m8 115 7 0 30.43 23.35Kitgum Patongo 3 1988m6 2005m12 211 3 0 14.33 11.02Lira Dokolo North 42 2003m8 2004m2 7 2 40 0.33 1.51Lira Erute South and Lira Mn 118 1987m3 2005m4 218 118 0 23.34 14.87Lira Kioga East 42 2004m1 2004m2 2 5 37 0.21 0.61Lira Moroto County East 78 1988m6 2004m1 188 78 0 25.91 18.29Lira Otuke East 25 1998m5 2005m10 90 13 12 12.04 12.93Soroti Kaberamaido County 39 1987m8 2003m11 196 39 0 23.64 15.22Soroti Kalaki County 1 1987m8 2003m11 196 1 0 16.00 .Soroti Orungo 4 2003m3 2005m9 31 4 0 15.75 14.08Soroti Soroti County North 2 2003m6 2003m12 7 2 0 7.00 0.00Soroti Soroti County South/Soroti 64 1987m3 2005m9 223 64 0 24.05 14.20Soroti Usuk County West 4 1998m4 2004m11 80 4 0 26.50 14.84Arua Kei 41 1998m7 1998m7 1 9 32 0.22 0.42Moyo West Moyo County East 2 2001m5 2005m4 48 2 0 19.00 19.80Moyo West Moyo County West 18 1996m9 2001m3 55 10 8 17.67 19.83Nebbi Jonam South 19 1997m10 1997m10 1 8 11 0.42 0.51Nebbi Padyere Central 38 1996m9 2006m8 120 26 12 19.32 19.81Kotido Abim 3 2003m5 2003m5 1 3 0 1.00 0.00Kotido Kaabong 5 2004m6 2004m6 1 5 0 1.00 0.00Moroto Matheniko South/Moroto Mn 21 1998m12 2004m2 63 9 12 5.48 8.65Jinja Jinja District 154 1987m10 2003m6 189 154 0 24.65 15.60Kamuli Budiope West 65 2004m2 2004m2 1 5 60 0.08 0.27Kapchorwa Binyini 2 2003m3 2003m4 2 2 0 2.00 0.00Hoima Bugahya North 24 1995m10 1996m7 10 1 23 0.21 1.02Masindi Budongo 26 2001m3 2005m12 58 3 23 4.77 13.80Masindi Kiryandongo 4 2005m11 2005m11 1 4 0 1.00 0.00Mubende Buwekula North 46 2000m9 2000m9 1 28 18 0.61 0.49Bushenyi Katerera 34 2000m12 2000m12 1 17 17 0.50 0.51

40

Appendix 3

As discussed in Section 3.3, the majority of LRA conflict events occurred in districts

dominated by Nilotic ethnic groups, with the former districts of Gulu, Kitgum, Lira,

Apac and Soroti accounting for 1842 of the 1947 recorded battles or acts of one-sided

violence listed in the ACLED dataset (94.6%). The ethnic Acholi (Gulu and Kitgum),

Langi (Lira and Apac) and Iteso (Soroti), within these areas, share anthropological links

to some of the tallest people in Africa, notably the Dinka of southern Sudan, and the

Maasai (Samburu) and Turkana, of neighbouring Kenya (Shoup, 2011).

To establish whether significant height differences exist between the adult women

of these Nilotic groups and the predominantly Bantu African groups further South, a

sample of 1720 adult women is taken from the 1995 DHS survey. This sample only

includes women who were at least 18 years old (to 49 years old) in 1986 and, therefore,

should have achieved full adult height before the LRA conflict began.35

The first model in Table A3 reports results from a simple bivariate regression of

women’s ethnicity on height, where the dependent variable is a binary indicator of

belonging to one of the three ethnic groups (Acholi, Langi or Iteso). The second es-

timation reports similar results, disaggregating the groups to establish whether height

disparities are apparent in all three ethnicities (relative to those not in these groups).

Based on this simple analysis, the average height advantage of these women appears

to be upward of 4.7 centimetres, with significant results for all groups individually.

Furthermore, controlling for variations in the ages of women, in models 3 and 4, does

nothing to change this interpretation. The results in models 5 to 8 repeat the estimations

35Unfortunately, the 1995 survey is the only relevant DHS wave where ethnicity is recorded. How-ever, the dominance of the Nilotic groups, in the districts listed, is supported by the self-reported ethnicaffiliation of the sample of women from this survey. In Gulu, 94.87% of the women identified as beinga member of the Acholi group, in Lira and Apac, 89.5% and 93.6% identified as Langi, while the Itesoethnic group accounted for 73.3% in Soroti district.

41

for the sub-sample of women coming only from conflict localities. The similarities in

the magnitude of the coefficients strongly suggests that the disparity in height between

these ethnic groups and others is also present within conflict localities.

Table A3: Ethnicity as a Determinant of Height - Women 18+ in 1986

Dep. Variable:height (cms) (1) (2) (3) (4) (5) (6) (7) (8)

Acholi/Langi/Iteso 4.757*** 4.817*** 4.282*** 4.538***(0.415) (0.418) (0.617) (0.653)

Acholi 4.012*** 4.173*** 3.460*** 3.895***(0.539) (0.575) (0.725) (0.888)

Langi 5.478*** 5.466*** 4.722*** 5.005***(0.518) (0.545) (0.696) (0.716)

Iteso 4.357*** 4.450*** 3.894*** 4.051***(0.554) (0.563) (0.812) (0.845)

Observations 1720 1720 1720 1720 295 295 295 295R2 0.070 0.071 0.083 0.084 0.101 0.104 0.173 0.176Age Fixed Effects No No Yes Yes No No Yes Yes

Standard errors in parentheses * p<0.1, ** p<0.05, *** p<0.01

42

Appendix 4

Table A4: Table 2 Results - Control Coefficients Reported

Dep. Variable:height-for-age z-score (1) (2) (3) (4) (5) (6)

Conf Locality * Exposed -0.239 -0.369** -0.388**(0.187) (0.186) (0.195)

Conf Locality * Ex Duration 0.001 -0.011** -0.011**(0.003) (0.005) (0.005)

Female child 0.146*** 0.149*** 0.152*** 0.147*** 0.148*** 0.151***(0.031) (0.032) (0.031) (0.031) (0.032) (0.031)

Birth order 4-6 -0.101** -0.102**(0.046) (0.046)

Birth order 7+ -0.296*** -0.296***(0.065) (0.066)

Mother’s height (cms) 0.049*** 0.049***(0.003) (0.003)

Mother married/cohabiting 0.036 0.037(0.067) (0.067)

Mother’s age (years) 0.017*** 0.017***(0.004) (0.004)

Mother primary educated 0.053 0.053(0.046) (0.046)

Mother secondary educated 0.251*** 0.252***(0.060) (0.060)

Mother is daughter of head -0.041 -0.040(0.101) (0.101)

Mother not head/daughter -0.240** -0.238**(0.104) (0.106)

Mother’s time in location -0.001 -0.001(0.002) (0.002)

Urban DHS cluster 0.424*** 0.421***(0.068) (0.068)

Head’s age (years) 0.005** 0.005**(0.002) (0.002)

Head is male 0.057 0.055(0.051) (0.051)

Observations 9202 9202 9202 9202 9202 9202R2 0.079 0.147 0.205 0.079 0.147 0.205Locality Fixed Effect Yes Yes Yes Yes Yes YesBirth Cohort Fixed Effect Yes Yes Yes Yes Yes YesDistrict-Specific Trend No Yes Yes No Yes Yes

Locality clustered standard errors in parentheses * p<0.1, ** p<0.05, *** p<0.01

43

Table 2.A4 Continued

Dep. Variable:height-for-age z-score (7) (8) (9) (10) (11) (12)

Conf Locality * Ex 1-6m -0.191 -0.200 -0.234(0.213) (0.215) (0.218)

Conf Locality * Ex > 6m -0.315* -0.659*** -0.651***(0.165) (0.194) (0.222)

Conf Locality * Ex 1-12m -0.231 -0.298 -0.317(0.197) (0.201) (0.204)

Conf Locality * Ex > 12m -0.258 -0.566*** -0.585**(0.171) (0.200) (0.227)

Female child 0.146*** 0.149*** 0.153*** 0.146*** 0.148*** 0.152***(0.031) (0.032) (0.031) (0.031) (0.032) (0.031)

Birth order 4-6 -0.101** -0.102**(0.046) (0.046)

Birth order 7+ -0.295*** -0.295***(0.066) (0.066)

Mother’s height (cms) 0.049*** 0.049***(0.003) (0.003)

Mother married/cohabiting 0.037 0.036(0.067) (0.067)

Mother’s age (years) 0.017*** 0.017***(0.004) (0.004)

Mother primary educated 0.053 0.054(0.046) (0.046)

Mother secondary educated 0.250*** 0.253***(0.060) (0.060)

Mother is daughter of head -0.040 -0.040(0.101) (0.101)

Mother not head/daughter -0.239** -0.241**(0.105) (0.105)

Mother’s time in location -0.001 -0.001(0.002) (0.002)

Urban DHS cluster 0.423*** 0.422***(0.069) (0.068)

Head’s age (years) 0.005** 0.005**(0.002) (0.002)

Head is male 0.057 0.056(0.051) (0.051)

Observations 9202 9202 9202 9202 9202 9202R2 0.079 0.148 0.206 0.079 0.147 0.205Locality Fixed Effect Yes Yes Yes Yes Yes YesBirth Cohort Fixed Effect Yes Yes Yes Yes Yes YesDistrict-Specific Trend No Yes Yes No Yes Yes

Locality clustered standard errors in parentheses * p<0.1, ** p<0.05, *** p<0.01

44

Table A4 reports identical results to those found in Table 2, with the inclusion of the

coefficients on the control variables (listed beneath Table 2). Focusing on controls at the

level of the child, significant differences in height-for-age z-scores are notable between

children’s gender groups, although, this is likely to only represent the extent to which

Ugandan children (of either gender) fit the growth patterns of the reference population.

The variables representing birth order clearly influences the child’s height, with those in

the first three positions having significant height advantages over sibling born later. The

characteristics of the child’s mother also clearly exert influence on the results, while

both the age of the head and the mother appear positively correlated with childhood

height.36 Interestingly, the relationship between these two household members shows

evidence of a relative height deficit for children who are neither, the child or grandchild

of the head (However, this does not necessarily imply a causal effect). Unsurprisingly,

children in urban locations (as defined by the original DHS clusters) are taller than

their rural counterparts, on average, yet there is no evidence of any overall effect on

children’s height associated with the gender of the household head, the marital status of

the child’s mother or the time she has spent in the current DHS cluster location.

36The head of the household and the mother of the child are different household members for 86.4%of the observations within the sample used in Table 2.

45

Appendix 5

Table A5 summarises the migration data taken from the Ugandan National Household

Survey 2005/06.37 Of those 669 households who migrated from one of the affected

districts (listed in Table A5), only 147 left the original district (21.4%). Furthermore,

within the group of 283 households who relocated specifically to avoid insecurity, only

30 (10.6%) moved to a new district.

Table A5: Uganda National Household Survey 2005/06 Migration History

Migrated Since 2001: For Any Reason To Avoid InsecuritySample Same Same New Same Same New

District From Total Total District Region Region Total District Region Region

Mubende 289 44 39 1 4 0 0 0 0Jinja 164 46 20 10 16 0 0 0 0Kamuli 320 37 23 12 2 0 0 0 0Kapchorwa 39 4 3 1 0 0 0 0 0Soroti 141 39 27 4 8 15 10 1 4Kotido 68 3 1 0 2 1 0 0 1Moroto 57 5 2 0 3 3 1 0 2Apac 316 61 38 22 1 11 5 6 0Gulu 213 65 58 3 4 46 43 2 1Kitgum 166 129 121 3 5 118 115 1 2Lira 240 135 122 11 2 82 72 10 0Arua 367 34 31 1 2 0 0 0 0Moyo 57 16 15 1 0 7 7 0 0Nebbi 116 17 12 0 5 0 0 0 0Hoima 86 11 5 4 2 0 0 0 0Masindi 110 23 9 0 14 0 0 0 0

Total 2749 669 526 73 70 283 253 20 10

37Uganda Bureau of Statistics (2008), Uganda National Household Survey 2005/2006, Version 1.0 ofthe public use dataset, provided by the National Data Archive. www.ubos.org

46

Appendix 6

Table A6: Mother’s Time in Location on Height-for-Age, by District

District Obs. dy/dx se t stat p-value

Apac 337 −0.001 0.014 −0.094 0.926Arua 536 −0.004 0.004 −0.896 0.373Bushenyi 378 −0.002 0.005 −0.442 0.660Gulu 102 −0.215 0.020 −10.578 0.000***Hoima 141 −0.033 0.017 −1.999 0.049**Kamuli 651 0.000 0.010 0.010 0.992Kapchorwa 85 −0.001 0.003 −0.248 0.805Kitgum 63 −0.005 0.008 −0.664 0.509Kotido 111 −0.031 0.011 −2.816 0.006***Lira 305 0.000 0.003 −0.048 0.962Masindi 117 −0.006 0.005 −1.054 0.295Moroto 115 −0.007 0.011 −0.623 0.535Moyo 119 0.000 0.005 0.082 0.935Mubende 398 0.001 0.008 0.191 0.849Nebbi 139 0.013 0.007 1.963 0.053*Soroti 224 0.004 0.002 1.618 0.110

Total 3975

Locality clustered standard errors reported * p<0.1, ** p<0.05, *** p<0.01

The results shown in Table A6 report the marginal effects and standard errors of an

additional year in the current location (based on the mother’s time in location) on height-

for-age z-score, by district. These marginal effects are based on an estimation of the

model shown below.

hazidt = δt +Districtd + θ1(Time in Locationi)

+θ1(Districtd ∗ Time in Locationi) + θXidt + υidt

(a1)

In model (a1) hazidt measure the height-for-age z-score of a child who resides in dis-

trict d. The variableDistrict represents a district fixed effect, while Time in Location

measures the number of years the child’s mother has lived in the current DHS cluster.

47

The termX represents the same control variables used in generating the results in Table

2. As the majority of relocation appears to have occurred within districts (see Appendix

5), only observations taken from children in the conflict-affected districts, who lived in

localities which had not yet been affected by the insurgency, are used in the estimations.

The results indicate the presence of a negative correlation between time in location

and height in observations from pre-conflict localities in Gulu, Hoima and Kotido dis-

tricts. This negative correlation could represent a selection effect, driven by a relocation

away from more severely affected localities, amongst relatively healthier children (see

discussion in Section 6.1).

48